|
|
[ Home | Optimize | Current | Project | Tools | Made in France | F.A.Q | Forums | Links | Sources Codes ]
There are many kinds of optimisations as there many way to do them. REMOVING THE MUL INSTRUCTION You can replace the multiply instruction by a left shit if you are using a power of two constant. You can make many left shifts for replacing a multiply instruction with any constant. Now here are some examples. Suppose you want to multiply EAX by 2, just write SHL EAX,1 Now suppose you need to multiply the same register by 100. For this operation it is very useful to get all the powers of two that we find in 100. This number is equal to 64 + 32 + 4 = 26 + 25 + 22. You need to use three registers for this operation. At the beginnig EAX contains the number to multiply. Initialize EBX and ECX with EAX. EAX = EBX = ECX = 100. Now write : SHL EAX,6 ; EAX *= 64 ADD EAX,EBX ; EAX = (X * 64) + (X * 32)
=> EAX = X * 96 Suppose that the registers are initialized with the value 97. This method is not the best. There are too many instructions. Just before the first SHL instruction you have two MOV instructions for setting the EAX register into EBX and ECX. The best way is given by Paul Hsieh. My method is only for learning purpose. Paul Hsieh
replaces these SHIFTS / ADD with the following instructions. Take a look at
Paul Hsieh's multiplication table. I am trying to decrypt it. In this example
EAX is initialized with the number to multiply. We suppose that EAX is now 97. The first line multiplies the EBX register by 3. The second multiplies EAX by 4. The third line multiplies by 32 the register EBX. The last line creates the result by adding EBX to EAX. This code is very compact. It shows us how to multiply a number by 3 with one line of code only. All the instructions take one clock. There are 4 lines and 4 * 1 != 3. How does it compute ? The two LEA instructions are executed at the same time and the result is one clock. The two others lines are not pairables so they are executed one by one, in that case they use two clocks. 1 + 2 = 3 clocks. REMOVE THE DIV INSTRUCTION Replacing the DIV instruction is more complex than replacing the MUL instruction. If we divide by a power of two, DIV can be replaced by a right shift. If we divide by an other constant, it can be quicker to apply the following formula : A = B / C = B * 1/C. The 1/C is equal to NOT C. For some divisions we can use a magic number. I had a post on MASM Forum for dividing by 100, it is at http://www.masm32.com/board/index.php?topic=5821.0 Use "Magic divider" (coded by The Svin) to generate other divisors
Many kinds of optimizations can be found Agner Fog's Web Site. My Own Methods What I suggest is to replace the INVOKE instruction when two or more calls are following themselves. I just want to say if you have the following code : INVOKE TranslateMessage,ADDR Msg It is tranlated like this : LEA EAX,Msg One way to optimize could be by pushing twice the EAX register like this the second LEA EAX,Msg and PUSH EAX are simply replaced with only one PUSH EAX. The new code is : LEA EAX,MSG What have we won ? One line, and one or two clocks ? Wonderful, but
there is better to do. We will keep the manner the parameters are pushed onto
the stack. When the call to the TranslateMessage is made, it pushes the return
address which is the address of the CALL DispatchMessage, and it jumps to the
TranslateMessage function. OK ? LEA EAX,Msg
; Load Structure address into EAX Is it clear ? Good, now there is a last optimizations to do. If the last call is followed by a JMP instruction. Wa have something like this : LEA EAX,Msg
; Load Structure address into EAX Now we just have to replace "PUSH OFFSET @F" with "PUSH OFFSET Loop". The final code is : @Loop
:
|
|
[ Home | Optimize | Current | Project | Tools | Made in France | F.A.Q | Forums | Links | Sources Codes ]
|
